Recurrent Multiscale Feature Modulation for Geometry Consistent Depth Learning.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• U-Net-like coarse-to-fine networks dominate dense prediction but have limitations.
• These networks suffer from error propagation and require heavy backbones.

Purpose of the Study:

• To propose a novel, lightweight network design for self-supervised monocular depth estimation.
• To overcome the limitations of existing coarse-to-fine network architectures.

Main Methods:

• Introduced Recurrent Multiscale Feature Modulation (R-MSFM), a lightweight network.
• R-MSFM uses per-pixel features, multiscale feature modulation, and recurrent refinement at fixed resolution.
• Developed a mask geometry consistency loss for geometry-aware depth learning.

Main Results:

• R-MSFM demonstrates superior performance in model size and inference speed.
• Achieved state-of-the-art results on KITTI and Make3D datasets.
• The proposed network design fundamentally avoids error propagation inherent in coarse-to-fine approaches.

Conclusions:

• R-MSFM presents an effective and efficient alternative for monocular depth estimation.
• The network's design and novel loss function contribute to improved performance and consistency.
• This work advances lightweight architectures for dense prediction tasks.